Uncertainty in the regressors in model.predict()

[ezekielbarnett]

trafficstars

Is your feature request related to a problem? Please describe. Using orbit to do forecasting, it is often the case that we have some uncertainty in our projections for regressors:

More specifically, in proj = model.predict(df=df_forecast), df_forecast consists of forecasts of the regressors, i.e. there is (probably) uncertainty about them, but currently uncertainty in the projections for proj only comes from parameter uncertainty, not "data uncertainty".

Describe the solution you'd like Take advantage of pyro to adapt how projections/projection uncertainty are generated. Certainly the simplest would be that projections in orbit are distributions, perhaps passing functions for the regressors? This flows more naturally in a multivariate forecasting scheme, but may be more challenging within the current framework of univariate / given regressors.

Describe alternatives you've considered Building up a multivariate forecasting scheme using pyro.

[edwinnglabs]

@ezekielbarnett brilliant idea. We have something similar to what you've described in our roadmap actually.